Anthropic·released 2025-05-221 source
Claude Sonnet 4 benchmark scores: 23 benchmarks tracked. 22% of its scores are independently reproduced — the rest are self-reported or unverified.
Leads on 0 of 23 benchmarks·best result 84.37 on MATH level 5 (reproduced)·5 of 23 independently reproduced·$3.3/$16.5 per M tokens
Consensus: LiteLLM · See every model’s pricing →
| Benchmark | Score | Evidence status | Measured |
|---|---|---|---|
| MATH level 5 | 84.37accuracy (%) | reproduced· optimizedT1 | 2025-05-22Epoch AI |
| Lech Mazur Writing | 81.4accuracy (%) | unverifiedT1 | 2025-05-22lechmazur/writing Github repository |
| GPQA diamond | 72.25accuracy (%) | reproduced· optimizedT1 | 2025-05-22Epoch AI |
| OTIS Mock AIME 2024-2025 | 71.08accuracy (%) | reproduced· optimizedT1 | 2025-05-22Epoch AI |
| METR Time Horizons | 61.97accuracy (%) | unverifiedT1 | 2025-05-22METR - Measuring AI Ability to Complete Long Tasks |
| DTBench | 61.78accuracy (%) | unverifiedT1 | 2025-05-22https://conceptualreasoning.ai/dtbench |
| Aider polyglot | 61.3accuracy (%) | unverified· optimizedT1 | 2025-05-22https://aider.chat/docs/leaderboards/ |
| Fiction.LiveBench | 46.9accuracy (%) | unverifiedT1 | 2025-05-22Fiction.live leaderboard |
| DeepResearch Bench | 46.6accuracy (%) | unverifiedT1 | 2025-05-22https://evals.futuresearch.ai/#drb |
| WeirdML | 46.11accuracy (%) | unverifiedT1 | 2025-05-22https://htihle.github.io/weirdml.html |
| OSWorld | 43.9accuracy (%) | unverified· optimizedT1 | 2025-05-22OS World Website |
| ARC-AGI | 40accuracy (%) | unverified· optimizedT1 | 2025-05-22ARC Prize Leaderboard |
| GeoBench | 37accuracy (%) | unverifiedT1 | 2025-05-22GeoBench leaderboard |
| Cybench | 35accuracy (%) | unverified· optimizedT1 | 2025-05-22Cybench leaderboard |
| SimpleBench | 34.6accuracy (%) | unverifiedT1 | 2025-05-22SimpleBench Leaderboard |
| LMCA | 34.12accuracy (%) | unverifiedT1 | 2025-05-22https://conceptualreasoning.ai/lmca |
| The Agent Company | 33.1accuracy (%) | unverifiedT1 | 2025-05-22TheAgentCompany leaderboard |
| FrontierMath-2025-02-28-Private | 7.26accuracy (%) | reproduced· optimizedT1 | 2025-05-22Epoch AI |
| ARC-AGI-2 | 5.93accuracy (%) | unverifiedT2 | 2025-05-22 |
| GSO-Bench | 4.9accuracy (%) | unverifiedT1 | 2025-05-22GSO Leaderboard |
| HLE | 3.11accuracy (%) | unverifiedT2 | 2025-05-22 |
| VPCT | 1accuracy (%) | unverifiedT1 | 2025-05-22VPCT leaderboard |
| FrontierMath-Tier-4-2025-07-01-Private | 0accuracy (%) | reproducedT1 | 2025-05-22Epoch AI |
Claims drawn from cited facts, not live model generation.
This model originates from Anthropic, the developer behind it. It dates to a 2025 release, placing it in a recent vintage rather than an older generation.
3 cited facts
This model is scored on 23 tracked benchmarks, but it currently tops none of them. Its record has been independently reproduced, meaning outside evaluation backs the scores up.
3 cited facts
Its mean gap to the best recorded score across its tracked benchmarks is 41.38 points.
1 cited fact
5 facts cross-checked across data sources: 5 single-source.
Source facts, citations, and refresh stamp for this record.
Sources: epoch_benchmarks, litellm_prices
Claude Sonnet 4 is an AI model developed by Anthropic, released 2025-05-22. Its benchmark record below tags every score as independently reproduced, self-reported, or unverified.
Claude Sonnet 4 has recorded scores on 23 benchmarks, each shown with its evidence status.
Claude Sonnet 4 has recorded scores on 23 benchmarks — Lech Mazur Writing, GPQA diamond, MATH level 5, OTIS Mock AIME 2024-2025, DTBench, WeirdML, and 17 more. The full table above shows each score with its evidence status.
5 of 23 recorded scores (22%) are independently reproduced rather than self-reported by the lab.
Claude Sonnet 4 has 23 tracked claims: 5 independently reproduced, 18 unverified.
Listed API pricing: $3.3 per million input tokens, $16.5 per million output tokens. See the pricing block for the full breakdown.